Showing posts with label fairness. Show all posts
Showing posts with label fairness. Show all posts

Friday, December 1, 2023

Fairness in algorithms: Hans Sigrist Prize to Aaron Roth

 The University of Bern's Hans Sigrist Prize has been awarded to Penn computer scientist Aaron Roth, and will be celebrated today.

Here are today's symposium details and schedule:

Here's an interview:

Aaron Roth: Pioneer of fair algorithms  In December 2023, the most highly endowed prize of the University of Bern will go to the US computer scientist Aaron Roth. His research aims to incorporate social norms into algorithms and to better protect privacy.  by Ivo Schmucki 

"There are researchers who sit down and take on long-standing problems and just solve them, but I am not smart enough to do that," says Aaron Roth. "So, I have to be the other kind of researcher. I try to define a new problem that no one has worked on yet but that might be interesting."

"Aaron Roth's own modesty may stand in the way of understanding the depth of his contributions. In fact, when he authored his doctoral thesis on differential privacy about 15 years ago and then wrote on the fairness of algorithms a few years later, terms like “Artificial Intelligence” and “Machine Learning” were far from being as firmly anchored in our everyday lives as they are today. Aaron Roth was thus a pioneer, laying the foundation for a new branch of research.

"I am interested in real problems. Issues like data protection are becoming increasingly important as more and more data is generated and collected about all of us," says Aaron Roth about his research during the Hans Sigrist Foundation’s traditional interview with the prize winner. He focuses on algorithmic fairness, differential privacy, and their applications in machine learning and data analysis.

...

"It is important that more attention is paid to these topics," says Mathematics Professor Christiane Tretter, chair of this year's Hans Sigrist Prize Committee. Tretter says that many people perceive fairness and algorithms as two completely different poles, situated in different disciplines and incompatible with each other. "It is fascinating that Aaron Roth’s work shows that this is not a contradiction."

...

"The first step to improving the analysis of large data sets is to be aware of the problem: "We need to realize that data analysis can be problematic. Once we agree on this, we can consider how we can solve the problems," says Aaron Roth."





Monday, October 23, 2023

Waitlist equity, when not everyone can wait a long time, by Afshin Nikzad and Philipp Strack

 Patients waiting for deceased donor kidneys are given priority in part by how long they have been on dialysis, while patients waiting for livers are prioritized according to how sick they are, sickest first.  When the wait is long, not everyone has an equal chance of surviving long enough to receive an organ. Here's a paper that suggests that service in random order (SIRO) has desirable equity properties. Efficiency depends on how patients' welfare and future prospects change while they wait.

Equity and Efficiency in Dynamic Matching: Extreme Waitlist Policies, by Afshin Nikzad and Philipp Strack, Management Science, forthcoming, Published Online:3 Oct 2023https://doi.org/10.1287/mnsc.2023.01212

Abstract: Waitlists are commonly used to allocate scarce resources, such as public housing or organs. Waitlist policies attempt to prioritize agents who wait longer by assigning them priority points (à la first come, first served). We show that such point systems can lead to severe inequality across the agents’ assignment probabilities unless they use randomization. In particular, deterministic point systems lead to a more unequal allocation than any other rule that prioritizes earlier arrivals, an axiom that ensures that agents who wait longer are treated (weakly) better. Among the policies abiding by this axiom, we show that service in random order (SIRO) leads to the most equal allocation. From a utilitarian perspective, we show that the planner faces no trade-off between equity and efficiency when the flow utility from waiting is nonnegative or negative and increasing over time. In these cases, SIRO is also the most efficient policy. However, when the flow cost of waiting increases over time, then the planner may face an efficiency–equity trade-off: SIRO remains the most equitable policy but may not be the most efficient one.


1. Introduction: Waitlists are a common way to allocate scarce resources, such as public housing,1 organs,2 or services such as call center support.3 There are many ways to decide who among the waiting agents receives an object once it becomes available. Some waitlists operate in a service-in-random-order (SIRO) manner and use lotteries to allocate objects to waiting agents, such as in the Diversity Immigrant Visa Program in the United States4 or Beijing’s license plate allocation.5 Many other waitlists follow designs akin to first come, first served (FCFS), in which whoever has waited for the longest time receives (priority points for) an object. For example, in the national kidney transplant waitlist in the United States, enrolled patients earn priority points for each day that they remain on the waitlist.6 Such rules ensure that an agent who waits longer is not treated worse than an agent with a shorter waiting time and otherwise identical characteristics.

"Prioritizing agents with longer waiting times, however, has a drawback: it implies that an agent with a longer lifetime, that is, an agent who can wait longer for an object, has a higher probability of receiving an object. This naturally leads to inequality in assignment probabilities across agents with varying lifetimes. For example, a first-come, first-served list would lead to many of the sickest patients never receiving an organ as they depart the system before having waited long enough to receive an organ. Such equity concerns, for example, play an important role in the context of organ allocations (Organ Procurement and Transplantation Network 2015). The high-level question we ask here is, what policy induces the least inequality among policies that give priority to agents who arrive earlier? Furthermore, is minimizing inequality aligned with the objective of a planner who maximizes the average of the agents’ utilities, or are there efficiency–equity trade-offs to be considered here?"

Monday, September 4, 2023

Covid medication: allocation, information, hesitancy, and uptake: what are some things we have learned?

 I've posted before about how informational advertising about vaccine availability and safety seems to have had a positive effect on vaccination rates among disadvantaged populations. There was particular concern in the U.S. at one point that Black people were less likely to receive vaccines and other medications than other Americans.

Today's post collects several papers about the effect of randomly allocating invitations for temporarily scarce Covid medications, while giving members of disadvantaged groups a higher probability of receiving an invitation.  Included will be an editorial warning us that we shouldn't be satisfied to judge the outcome of a market design by its intended outcome ("Moving Beyond Intent and Realizing Health Equity").

There are market design lessons in these last few years of Covid experience that I hope will help make the responses to future pandemics more effective. Not least of these is that the allocation of public health  and medical resources turns out to be quite different from  the allocation of other kinds of resources, in many important ways that reflect the broader economic and social environments in which different kinds of allocation takes place.

###

Here's a paper in the most recent issue of JAMA Health Forum, by a team that includes both medical professionals and market designers.

Weighted Lottery to Equitably Allocate Scarce Supply of COVID-19 Monoclonal Antibody , by Erin K. McCreary, PharmD1; Utibe R. Essien, MD, MPH2,3; Chung-Chou H. Chang, PhD4,5; Rachel A. Butler, MHA, MPH6; Parag Pathak, PhD7; Tayfun Sönmez, PhD8; M. Utku Ünver, PhD8; Ashley Steiner, BS9; Maddie Chrisman, PT, DPT10; Derek C. Angus, MD, MPH11; Douglas B. White, MD, MAS11, JAMA Health Forum. 2023;4(9):e232774. Sept. 1, doi:10.1001/jamahealthforum.2023.2774 

"Objective  To describe the development and use of a weighted lottery to allocate a scarce supply of tixagevimab with cilgavimab as preexposure prophylaxis to COVID-19 for immunocompromised individuals and examine whether this promoted equitable allocation to disadvantaged populations.

"Design, Setting, and Participants  This quality improvement study analyzed a weighted lottery process from December 8, 2021, to February 23, 2022, that assigned twice the odds of drug allocation of 450 tixagevimab with cilgavimab doses to individuals residing in highly disadvantaged neighborhoods according to the US Area Deprivation Index (ADI) in a 35-hospital system in Pennsylvania, New York, and Maryland. In all, 10 834 individuals were eligible for the lottery. Weighted lottery results were compared with 10 000 simulated unweighted lotteries in the same cohort performed after drug allocation occurred.

"Main Outcomes:  Proportion of individuals from disadvantaged neighborhoods and Black individuals who were allocated and received tixagevimab with cilgavimab.

"Results:  Of the 10 834 eligible individuals, 1800 (16.6%) were from disadvantaged neighborhoods and 767 (7.1%) were Black. Mean (SD) age was 62.9 (18.8) years, and 5471 (50.5%) were women. A higher proportion of individuals from disadvantaged neighborhoods was allocated the drug in the ADI-weighted lottery compared with the unweighted lottery (29.1% vs 16.6%; P < .001). The proportion of Black individuals allocated the drug was greater in the weighted lottery (9.1% vs 7.1%; P < .001). Among the 450 individuals allocated tixagevimab with cilgavimab in the ADI-weighted lottery, similar proportions of individuals from disadvantaged neighborhoods accepted the allocation and received the drug compared with those from other neighborhoods (27.5% vs 27.9%; P = .93). However, Black individuals allocated the drug were less likely to receive it compared with White individuals (3 of 41 [7.3%] vs 118 of 402 [29.4%]; P = .003).

...

"Conclusions and Relevance:  The findings of this quality improvement study suggest an ADI-weighted lottery process to allocate scarce resources is feasible in a large health system and resulted in more drug allocation to and receipt of drug by individuals who reside in disadvantaged neighborhoods. Although the ADI-weighted lottery also resulted in more drug allocation to Black individuals compared with an unweighted process, they were less likely to accept allocation and receive it compared with White individuals. Further strategies are needed to ensure that Black individuals receive scarce medications allocated."

...

"The lottery was repeated over several weeks, but we chose to examine only the first assignment. The interpretation of later rounds is problematic because eventually all individuals were offered tixagevimab with cilgavimab. By focusing on the first draw, we can specifically evaluate whether the intent of the lottery was met."

##############

Closely related reports:

White, D.B., McCreary, E.K., Chang, C.C.H., Schmidhofer, M., Bariola, J.R., Jonassaint, N.N., Persad, G., Truog, R.D., Pathak, P., Sonmez, T. and Unver, M.U., 2022. A multicenter weighted lottery to equitably allocate scarce COVID-19 therapeutics. American Journal of Respiratory and Critical Care Medicine, 206(4), pp.503-506.

Rubin, E., Dryden-Peterson, S.L., Hammond, S.P., Lennes, I., Letourneau, A.R., Pathak, P., Sonmez, T. and Ünver, M.U., 2021. A novel approach to equitable distribution of scarce therapeutics: institutional experience implementing a reserve system for allocation of COVID-19 monoclonal antibodies. Chest, 160(6), pp.2324-2331.*

White, D.B. and Angus, D.C., 2020. A proposed lottery system to allocate scarce COVID-19 medications: promoting fairness and generating knowledge. Jama, 324(4), pp.329-330.

###########

And here's an editorial in the same issue of JAMA Health Forum as the most recent article, pointing out that less-disadvantaged patients among those living in census blocks identified as disadvantaged (in particular  commercially insured and White patients) were much more likely to receive the treatment:

Moving Beyond Intent and Realizing Health Equity, by Atheendar S. Venkataramani, MD, PhD, Invited Commentary, September 1, 2023, JAMA Health Forum. 2023;4(9):e232525. doi:10.1001/jamahealthforum.2023.2525

"In a study published in this issue of JAMA Health Forum, McCreary and colleagues3 report on a landmark effort at the University of Pittsburgh Medical Center (UPMC) to distribute equitably a scarce monoclonal antibody resource, tixagevimab with cilgavimab, for COVID-19 preexposure prophylaxis in immunocompromised individuals. In December 2021, UPMC received an allotment of 450 doses of tixagevimab with cilgavimab from the Pennsylvania Department of Health to cover a large health system with 35 hospitals and 800 outpatient facilities through February 2022. In an ex ante effort to mitigate health disparities and respond to guidance from the Commonwealth of Pennsylvania to allocate scarce resources in a manner that accounts for multiple ethical objectives, UPMC convened an advisory group of clinicians, community stakeholders, and experts in community outreach.

...

"The lottery was constructed using the Area Deprivation Index (ADI) to ensure that patients in highly disadvantaged neighborhoods had an equal opportunity to access tixagevimab with cilgavimab. Patients living in neighborhoods with ADIs above a specific cutoff that has been shown to best target less affluent, rural, and Black patients received 2 entries in the lottery, compared with 1 entry for patients in more advantaged neighborhoods. In their study, McCreary and colleagues3 found that this process resulted in equitable access: similar proportions of individuals in more advantaged and more disadvantaged neighborhoods (about 28% in each group) received tixagevimab with cilgavimab during the study period, although Black patients who were allocated the drug in the lottery were significantly less likely to receive it compared with White patients (7.3% vs 29.4%).

...

"Having identified its patient population, UPMC required only patient addresses as well as publicly available data on ADIs to implement the lottery intervention. The ADIs are defined at the census block group level, which include about 1000 residents on average. Thus, UPMC was able to achieve equitable opportunity to access tixagevimab with cilgavimab across small localities with very different socioeconomic profiles.

...

On the other hand, higher-resolution data that specifically measure the types of intersecting, reinforcing, and cumulative disadvantages faced by historically marginalized groups5 may be needed to achieve equitable outcomes across other dimensions, such as race and ethnicity. Within census blocks, patients assigned the same ADI levels but who may have faced relatively fewer structural barriers compared with Black patients or patients receiving Medicaid—namely, commercially insured and White patients—were more likely to access tixagevimab with cilgavimab conditional on being allocated to receive it in the lottery

...

"The lower rates of drug receipt among Black patients also underscores the importance of complementary investments and operational decisions to address additional structural barriers to accessing medical technology.

...

"The study by McCreary and colleagues3 represents the type of courageous and rigorous work that is needed to chart a path forward in determining how best to bridge the access gap for leading-edge medical technology. Future work would benefit from the same type of clarity demonstrated in this study by including clear definitions for how equity should be operationalized, attempting to address fragmentation between clinical services and services that address social drivers of health, aligning incentives, and addressing historical barriers that have made it difficult to achieve health equity."

##########

*Earlier:

Saturday, August 14, 2021


Thursday, September 29, 2022

What is needed to gain support for effective algorithms in hiring, etc?

 Here's an experiment motivated in part by European regulations on transparency of algorithms.

Aversion to Hiring Algorithms: Transparency, Gender Profiling, and Self-Confidence  by Marie-Pierre Dargnies, Rustamdjan Hakimov and Dorothea Kübler

Abstract: "We run an online experiment to study the origins of algorithm aversion. Participants are either in the role of workers or of managers. Workers perform three real-effort tasks: task 1, task 2, and the job task which is a combination of tasks 1 and 2. They choose whether the hiring decision between themselves and another worker is made either by a participant in the role of a manager or by an algorithm. In a second set of experiments, managers choose whether they want to delegate their hiring decisions to the algorithm. In the baseline treatments, we observe that workers choose the manager more often than the algorithm, and managers also prefer to make the hiring decisions themselves rather than delegate them to the algorithm. When the algorithm does not use workers’ gender to predict their job task performance and workers know this, they choose the algorithm more often. Providing details on how the algorithm works does not increase the preference for the algorithm, neither for workers nor for managers. Providing feedback to managers about their performance in hiring the best workers increases their preference for the algorithm, as managers are, on average, overconfident."

"Our experiments are motivated by the recent debates in the EU over the legal requirements for algorithmic decisions. Paragraph 71 of the preamble to the General Data Protection Regulation (GDPR) requires data controllers to prevent discriminatory effects of algorithms processing sensitive personal data. Articles 13 and 14 of the GDPR state that, when profiling takes place, people have the right to “meaningful information about the logic involved” (Goodman and Flaxman 2017). While the GDPR led to some expected effects, e.g., privacy-oriented consumers opting out of the use of cookies (Aridor et al. 2020), the discussion over the transparency requirements and the constraints on profiling is still ongoing. Recently, the European Parliament came up with the Digital Services Act (DSA), which proposes further increasing the requirements for algorithm disclosure and which explicitly requires providing a profiling-free option to users, together with a complete ban on the profiling of minors. Our first treatment that focuses on the workers aims at identifying whether making the algorithm gender-blind and therefore unable to use gender to discriminate, as advised in the preamble of the GDPR and further strengthened in the proposed DSA, increases its acceptance by the workers. The second treatment is a direct test of the importance of the transparency of the algorithm for the workers. When the algorithm is made transparent in our setup, it becomes evident which gender is favored. This can impact algorithm aversion differently for women and men, for example if workers’ preferences are mainly driven by payoff maximization.

"The treatments focusing on the managers’ preferences aim at understanding why some firms are more reluctant than others to make use of hiring algorithms. One possible explanation for not adopting such algorithms is managerial overconfidence. Overconfidence is a common bias, and its effect on several economic behaviors has been demonstrated (Camerer et al. 1999, Dunning et al. 2004, Malmendier and Tate 2005, Dargnies et al. 2019). In our context, overconfidence is likely to induce managers to delegate the hiring decisions to the algorithm too seldom. Managers who believe they make better hiring decisions than they actually do, may prefer to make the hiring decisions themselves. Our paper will provide insights about the effect of overconfidence on the delegation of hiring decisions to algorithms. Similar to the treatments about the preferences of workers, we are also interested in the effect of the transparency of the algorithm on the managers’ willingness to delegate the hiring decisions. Disclosing the details of the algorithm can increase the managers’ trust in the algorithm."

Sunday, August 15, 2021

Fair algorithms for selecting citizen assemblies, in Nature

 Here's a paper that grapples with the problem that not every group in society is equally likely to accept an appointment for which they have been selected, which complicates the problem of selecting representative committees while also giving each potential member approximately the same chance of being selected.

Fair algorithms for selecting citizens’ assemblies. by Bailey Flanigan, Paul Gölz, Anupam Gupta, Brett Hennig & Ariel D. Procaccia, Nature (2021). https://doi.org/10.1038/s41586-021-03788-6

Abstract: Globally, there has been a recent surge in ‘citizens’ assemblies’1, which are a form of civic participation in which a panel of randomly selected constituents contributes to questions of policy. The random process for selecting this panel should satisfy two properties. First, it must produce a panel that is representative of the population. Second, in the spirit of democratic equality, individuals would ideally be selected to serve on this panel with equal probability2,3. However, in practice these desiderata are in tension owing to differential participation rates across subpopulations4,5. Here we apply ideas from fair division to develop selection algorithms that satisfy the two desiderata simultaneously to the greatest possible extent: our selection algorithms choose representative panels while selecting individuals with probabilities as close to equal as mathematically possible, for many metrics of ‘closeness to equality’. Our implementation of one such algorithm has already been used to select more than 40 citizens’ assemblies around the world. As we demonstrate using data from ten citizens’ assemblies, adopting our algorithm over a benchmark representing the previous state of the art leads to substantially fairer selection probabilities. By contributing a fairer, more principled and deployable algorithm, our work puts the practice of sortition on firmer foundations. Moreover, our work establishes citizens’ assemblies as a domain in which insights from the field of fair division can lead to high-impact applications.

...

"To our knowledge, all of the selection algorithms previously used in practice (Supplementary Information section 12) aim to satisfy one particular property, known as ‘descriptive representation’ (that the panel should reflect the composition of the population)16. Unfortunately, the pool from which the panel is chosen tends to be far from representative. Specifically, the pool tends to overrepresent groups with members who are on average more likely to accept an invitation to participate, such as the group ‘college graduates’.  

...

"Selection algorithms that pre-date this work focused only on satisfying quotas, leaving unaddressed a second property that is also central to sortition: that all individuals should have an equal chance of being chosen for the panel.

...

"Although it is generally impossible to achieve perfectly equal probabilities, the reasons to strive for equality also motivate a more gradual version of this goal: making probabilities as equal as possible, subject to the quotas. We refer to this goal as ‘maximal fairness’

...

"the algorithms in our framework (1) explicitly compute a maximally fair output distribution and then (2) sample from that distribution to select the final panel (Fig. 1). Crucially, the maximal fairness of the output distribution found in the first step makes our algorithms optimal. To see why, note that the behaviour of any selection algorithm on a given instance is described by some output distribution; thus, as our algorithm finds the fairest possible output distribution, it is always at least as fair as any other algorithm."



Saturday, August 14, 2021

A lottery for antibody treatment, with slots reserved for vulnerable patients

 It's always good to see a collaboration between physicians and economists on allocating scarce resources, and here's a case report of allocating monoclonal antibodies in Boston (with some resemblance to school choice), forthcoming in the journal CHEST.

A novel approach to equitable distribution of scarce therapeutics: institutional experience implementing a reserve system for allocation of Covid-19 monoclonal antibodies  Emily Rubin, MD JD MSHP, Scott L. Dryden-Peterson, MD, Sarah P. Hammond, MD, Inga Lennes, MD MBA MPH, Alyssa R. Letourneau, MD MPH, Parag Pathak, PhD, Tayfun Sonmez, PhD, M. Utku Ünver, PhD.

DOI: https://doi.org/10.1016/j.chest.2021.08.003, To appear in: CHEST

"Background. In fall 2020, the Food and Drug Administration issued emergency use authorization for monoclonal antibody therapies (mAbs) for outpatients with Covid-19.  The Commonwealth of Massachusetts issued guidance outlining the use of a reserve system with a lottery for allocation of mAbs in the event of scarcity that would prioritize socially vulnerable patients for 20% of the infusion slots. The Mass General Brigham (“MGB”) health system subsequently implemented such a reserve system.

"Research Question. Can a reserve system be successfully deployed in a large health system in a way that promotes equitable access to mAb therapy among socially vulnerable patients with Covid-19?

...

"ResultsNotwithstanding multiple operational challenges, the reserve system for allocation of mAb therapy worked as intended to enhance the number of socially vulnerable patients who were offered and received mAb therapy. A significantly higher proportion of patients offered mAb therapy were socially vulnerable (27.0%) than would have been the case if the infusion appointments had been allocated using a pure lottery system without a vulnerable reserve (19.8%) and a significantly higher proportion of patient who received infusions were socially vulnerable (25.3%) than would have been the case if the infusion appointments had been allocated using a pure lottery system (17.6%)

...

"The reserve for vulnerable patients was a “soft” reserve, meaning that if there were not enough patients in either the high SVI or high incidence town categories to fill the vulnerable slots, those slots were allocated to patients who were next in line by overall lottery number. This was done in order to avoid unused capacity for a therapy that is time sensitive and requires significant infrastructure to provide. Once the lottery had been run, dedicated, primarily multilingual clinicians who had been trained to discuss the therapies with patients called patients to verify eligibility and engage in a shared-decision making conversation to determine whether the patient would like to receive an infusion.

Early experience with running the lottery prior to patient engagement revealed that a large number of patients declined the therapy once offered, were deemed ineligible once contacted, or wished to discuss the therapy with a trusted clinician. The process subsequently was changed to allow clinicians to enter referrals for their own patients once they established patient interest (“manual referrals”). 

...

"All of the 274 patients who were guaranteed slots and 206 of 368 patients on the wait list were called, for a total of 480 patients called. The number of wait list patients called on a given day was a function of both how many of the guaranteed slots were not filled and how much capacity there was in the system to make phone calls on any given day. Of those patients who were called, 132 (27.5%) declined, 33 (6.9%) were deemed ineligible by virtue of being asymptomatic, 19 (4.0%) were deemed ineligible by virtue of having severe symptoms, 11 (2.3%) had been or were planning to be infused elsewhere, 61 (12.7%) could not be reached, and 191 were infused (39.8% of those called and 9.7% of total referred patients).

...

"Had we operated a pure lottery with no reserve for socially vulnerable patients, and all other factors had remained constant, 19.8% of patients offered therapy (88) would have been in the top SVI quartile as opposed to 27.0% (120) in our actual population, and 17.6% of infused patients (32) would have been in the top SVI quartile as opposed to 25.3% (46) in our actual population.

...

"The system we describe is to our knowledge the first instance of a reserve system being used to allocate scarce resources at the individual level during a pandemic.

"A reserve system with lottery for tiebreaking within categories can be straightforward to operate if there are few or no steps between the assignment of lottery spots and the distribution of the good. This could be true, for example, of allocation of antiviral medications to inpatients with Covid-19. In the case of monoclonal antibody therapies, there were multiple factors that could and often did interrupt the trajectory between allocation and distribution. These included the complexity of administering infusion therapy, the time sensitive nature of the therapy, the relative paucity of evidence for the therapy at the time the mAb program started, and the dynamic nature of Covid-19. The conversations with patients about a therapy that held promise but did not yet have strong evidence to support its efficacy and had not been formally FDA approved were often challenging and time consuming. Many patients identified for allocation were difficult or impossible to reach. Others declined therapy once it was offered and discussed, or had become either too well or too sick to be candidates for the therapy once they were reached.

...

"Notwithstanding significant challenges, the reserve system implemented in our health system for allocation of mAb therapy worked as intended to enhance the number of socially vulnerable patients who were offered the therapy. A significantly higher proportion of socially vulnerable patients were offered mAb therapy than would have been if the infusion appointments had been allocated using a pure lottery system without a vulnerable reserve. The intended enhancement of the pool of vulnerable patients who actually received monoclonal antibody therapy was counterbalanced to some extent by the disproportionate number of vulnerable patients who declined therapy, but even fewer socially vulnerable patients would have received the therapy if the lottery system had not included a vulnerable reserve. 

Wednesday, July 28, 2021

Redistribution through markets, in Econometrica, by Dworczak, Kominers, and Akbarpour

 Market designs involving taxes, or rationing, in the latest Econometrica, Vol. 89, No. 4 (July, 2021), 1665–1698:

REDISTRIBUTION THROUGH MARKETS by PIOTR DWORCZAK, SCOTT DUKE KOMINERS,  and MOHAMMAD AKBARPOUR

Abstract: "Policymakers frequently use price regulations as a response to inequality in the markets they control. In this paper, we examine the optimal structure of such policies from the perspective of mechanism design. We study a buyer-seller market in which agents have private information about both their valuations for an indivisible object and their marginal utilities for money. The planner seeks a mechanism that maximizes agents’ total utilities, subject to incentive and market-clearing constraints. We uncover the constrained Pareto frontier by identifying the optimal trade-off between allocative efficiency and redistribution. We find that competitive-equilibrium allocation is not always optimal. Instead, when there is inequality across sides of the market, the optimal design uses a tax-like mechanism, introducing a wedge between the buyer and seller prices, and redistributing the resulting surplus to the poorer side of the market via lump-sum payments. When there is significant same-side inequality that can be uncovered by market behavior, it may be optimal to impose price controls even though doing so induces rationing."

****************

" the classic idea that competitive-equilibrium pricing maximizes welfare relies on an implicit assumption that the designer places the same welfare weight on all agents in the market. Thus, the standard economic intuitions in support of competitive equilibrium pricing become unreliable as the dispersion of wealth in a society expands."

Thursday, March 4, 2021

Vaccine supply and delivery in Germany: I'm interviewed in Zeit

 Here's an interview in the German newspaper Zeit, in which I was asked in early February about the vaccine rollout here and there. (Google translate is pretty readable, although some of the Q&A is a bit garbled by the translation from English to German and re-translation back into English...)

"Die Welt kann es sich leisten, einiges zu bezahlen" Alvin Roth weiß, wie man begehrte Güter effizient verteilt. Er hat den Nobelpreis dafür bekommen. Ein Gespräch über knappen Impfstoff und wie er vermehrt werden kann.  Interview: Lisa Nienhaus

Google Translate: "The world can afford to pay a lot" Alvin Roth knows how to efficiently distribute desirable goods. He got the Nobel Prize for it. A conversation about scarce vaccine and how it can be propagated. Interview: Lisa Nienhaus February 15, 2021,

The interview starts off talking about congestion, and line jumping, and the tradeoffs between speed and fairness (and how it's really costly to allow some vaccine to expire unused in the name of fairness).  It then turns to shortages of vaccine in the near term:

ZEIT ONLINE: Attempts are being made to build new production facilities. But in Germany we are - to be honest - pretty late.

Roth: But now is not the time to give up. Everything we build now may help us in August. Even if Germany is running late, there is still time to expand production facilities. Especially since these systems would certainly not have to be destroyed after Covid. Being able to produce mRNA vaccines oneself is also a good thing in the future. Vaccine production is not that complicated. You can build production facilities anywhere. And you should too.

ZEIT ONLINE: It's not happening on a large scale yet. What to do?

Roth: Laws are really useful for that. Pfizer / BioNTech and Moderna could be forced to license the production technology to other German pharmaceutical companies.

ZEIT ONLINE: That sounds radical.

Roth: I only think it's logical. If you had a pharmaceutical company, you'd think, "I'm paid by the dose. I've got enough capacity to ship to the whole world in the next year and a half. Why should I hurry?" There is no need to set up production facilities just to supply the world in six months instead of 18. It doesn't make any difference from a business perspective. But for the German or American government, these two options are by no means equivalent. It is important that we vaccinate quickly. We need a lot more production capacity than the pharmaceutical companies think it makes sense.

ZEIT ONLINE: Economists rarely suggest such a strong market intervention. And that also applies to companies that we must first be grateful to because they show us a way out of lockdown.

Roth: It's a global pandemic. It is economically necessary to think about how to avert the damage to the economy. But of course you have to pay the manufacturers. Many forget that.

ZEIT ONLINE: How fair the companies think that probably depends on how much you pay them.

Roth: Yes. But the world can afford to pay a lot. Because the world economy is currently largely at a standstill. We have a multi-trillion dollar economy. Paying a billion to save a trillion is good business.

ZEIT ONLINE: Why is that not happening so far?

Roth: The pharmaceutical companies themselves don't think that way at the moment. But we need the vaccine now. And it's very expensive for the world to shut down its economy like that. If you lose a few percentage points of GDP growth in Germany, that's a huge number. And there is almost no amount to pay to license the vaccine that is not worth it.

Wednesday, March 3, 2021

Anger at vaccine line jumping

 There is some tension between getting populations vaccinated quickly and ensuring that priorities for who gets vaccinated first are carefully followed.  In some places we have seen the costs of adhering too strictly to priorities when enough high priority people are hard to find quickly.  In other places we see the costs of ignoring priorities.

Here's a NY Times story on corruption in South America (followed by a Guardian story about the difficulty of stopping tech-savvy Californians from grabbing appointments meant for underserved minorities):

‘V.I.P. Immunization’ for the Powerful and Their Cronies Rattles South America. A wave of corruption scandals is exposing how the powerful and well-connected in South America jumped the line to get vaccines early. Public dismay is turning into anger.   By Mitra Taj, Anatoly Kurmanaev, Manuela Andreoni and Daniel Politi

"The hope brought by the arrival of the first vaccines in South America is hardening into anger as inoculation campaigns have spiraled into scandal, cronyism and corruption, rocking national governments and sapping trust in the political establishment.

"Four ministers in Peru, Argentina and Ecuador have resigned this month or are being investigated on suspicion of receiving or providing preferential access to scarce coronavirus shots. Prosecutors in those countries, and in Brazil, are examining thousands more accusations of irregularities in inoculation drives, most of them involving local politicians and their families cutting in line.

...

“People find it much more difficult to tolerate corruption when health is at stake,” said Mariel Fornoni, a pollster in Buenos Aires.

The brazen nature of some of the scandals — which mirror similar affairs in LebanonSpain and the Philippines — has outraged the region.

...

"Earlier this month, the doctor conducting Peru’s first vaccine trial acknowledged inoculating nearly 250 politicians, notables and their relatives, as well as university administrators, interns and others, with undeclared extra doses. Some had received three doses, according to the trial’s director, Dr. Germán Málaga, in an attempt to maximize their immunity."

***************

And here's the Guardian, on California:


"Access codes meant to give Californians of color priority access to Covid-19 vaccine slots have been getting passed around among other residents in the state, allowing some to cut the line and get appointments meant for underserved Black and Latino residents.

Misuse of these codes was reported at vaccine sites in Los Angeles and the Bay Area, said Brian Ferguson, spokesperson for the California office of emergency services, to the Guardian.

"The codes were one of the tools devised by California leaders to address inequities in vaccine distribution in the state. They were given out to leaders and non-profits in the Black and Latino communities in LA and the Bay Area to administer to eligible individuals...

"Instead, the codes ended up passed on by text message and email, oftentimes with misinformation. “My daughter says that the Oakland Coliseum needs to fill up appointment slots in the next few days to prevent spoilage of excess vaccines!” read an email that Oakland resident Jhumpa Bhattacharya received from a friend on Monday. “If you are interested in getting a vaccine before this Wednesday, the link and access code are pasted below."
...
"State officials thought that by handing out vaccine access codes through community leaders, they would bridge any cultural or language barriers while also addressing the issue of the digital divide by giving these eligible individuals special access to the website, Ferguson said. “We don’t want people to be able to get appointments based on who has the fastest internet connection,” he said.

"Since learning of the misuse, the state will begin issuing individualized codes rather than group codes next week. In addition to these codes, the state has been setting up mobile vaccination clinics in these specific communities in hopes of reaching these underserved residents."

Friday, January 29, 2021

Vaccine delivery in the U.S. continues to be congested

As of today, congestion is still competing with short supply to limit vaccination in the US.

Some doses are being wasted or delayed in the name of fairness,  to better honor the priority orderings being used in each state, some doses are being sequestered for second vaccinations rather than being used now for first vaccinations, and some regions and/or providers have too little vaccine on hand, or too little predictability of supply to plan efficient distribution.

 USA today has the story:

Amid sputtering COVID-19 vaccine rollout, 16 states have used less than half of distributed doses  by Ken Alltucker

"The Biden administration has vowed drug companies will make enough vaccine to immunize 300 million Americans by the end of the summer.

But getting the vaccine from the factory to the arms of people has been anything but smooth. Of 47.2 million doses shipped to states and nursing homes, 24.6 million doses have been administered, the Centers for Disease Control and Prevention reported Thursday. 

The nation's slow rollout has boiled over from California, which tapped Blue Shield of California to allocate vaccines, to Maryland where Gov. Larry Hogan implored the federal government to send more doses of the potentially life-saving vaccine.

An Arlington, Virginia, hospital canceled 10,000 vaccine appointments, citing the state's decision to send doses to county health departments rather than directly to hospitals and other health providers.  In Minnesota, a vaccine lottery offered just 8,000 appointments to more than 226,000 people who signed up over a 24-hour period this week."

Saturday, January 2, 2021

Vaccine supply chain woes

Supply chains are boring, until things are in short supply. And there are many steps in a supply chain that can cause supplies to be short. Below are some news stories on how the U.S. is having trouble delivering vaccines, with the limiting factors not yet being shortage of the vaccines themselves.

I notice a few things about these stories. 

  • It seems to be widely recognized that it is worth spending billions (or at least hundreds of millions) to save trillions (i.e. to speed up vaccinations to hasten the reopening of the economy).
  • It seems also to be widely recognized that it would be regarded as repugnant to allocate initial inoculations by charging high prices for them while they are scarce: instead we are trying to establish priority orders for recipients: e.g. first health care workers and the elderly in nursing homes, then the independent elderly and the ill, etc,
  • Keeping strictly to priorities may partly be what is slowing down vaccinations: when not enough high priority people show up, the vaccines go back in the freezer to wait for the next day (at least I hope they go back in the freezer, and are not spoiled and unusable by the next day).  It might be better to try to find people ready to be vaccinated, when it's hard to find enough high priority people quickly.
  • A lack of confidence that more vaccine doses will be reliably arriving on schedule is causing some stockpiling, which is the enemy of fast distribution.
  • Holiday schedules make it hard to get lots of people vaccinated fast; maybe we'll do better this coming week.
Here's a story from the Financial Times:

Trump administration admits missing Covid vaccination goals--Officials say US states have used only a fifth of the doses they were given  by Kiran Stacey 

"Officials had aimed to distribute enough doses to vaccinate 20m people by the end of the year, but recently admitted they were not likely to hit that target until early January after underestimating how long it would take to perform quality control checks on manufactured doses.

"Figures released by the federal government, however, show a bigger hurdle is getting the vaccines to people once they have been manufactured and sent out. The US Centers for Disease Control and Prevention said on Wednesday just under 2.6m people in the country had been vaccinated, even though 12.4m doses had been distributed.

...

"Nancy Messonnier, the director of the National Center for Immunisation and Respiratory Diseases, blamed a range of factors. She said part of the problem was that pharmacies that were largely responsible for vaccinating people in care homes had been waiting to schedule appointments until they could be sure they had enough doses to perform booster shots."

************

From the NY Times:

Here’s Why Distribution of the Vaccine Is Taking Longer Than Expected--Health officials and hospitals are struggling with a lack of resources. Holiday staffing and saving doses for nursing homes are also contributing to delays.  By Rebecca Robbins, Frances Robles and Tim Arango

"In Florida, less than one-quarter of delivered coronavirus vaccines have been used, even as older people sat in lawn chairs all night waiting for their shots. In Puerto Rico, last week’s vaccine shipments did not arrive until the workers who would have administered them had left for the Christmas holiday. In California, doctors are worried about whether there will be enough hospital staff members to both administer vaccines and tend to the swelling number of Covid-19 patients.

"These sorts of logistical problems in clinics across the country have put the campaign to vaccinate the United States against Covid-19 far behind schedule in its third week, raising fears about how quickly the country will be able to tame the epidemic.

...

"Complicating matters, the county health department gets just a few days of notice each week of the timing of its vaccine shipments. When the latest batch arrived, Dr. Gayles’s team scrambled to contact people eligible for the vaccine and to set up clinics to give out the doses as fast as possible.

...

"In Florida, some hospital workers offered the vaccine declined it, and those doses are now designated for  other vulnerable groups like health care workers in the community and the elderly, but that rollout has not quite begun

...
"It may be more difficult, public health officials say, to vaccinate the next wave of people, which will most likely include many more older Americans as well as younger people with health problems and frontline workers. Among the fresh challenges: How will these people be scheduled for their vaccination appointments? How will they provide documentation that they have a medical condition or a job that makes them eligible to get vaccinated? And how will pharmacies ensure that people show up, and that they can do so safely?"

Sunday, June 28, 2020

Course allocation with minimum quotas, by Bichler, Hammerl, Waldherr, and Morrill

Here's a course allocation problem:, "Our specific task was to assign students to classes in the computer science and information systems department at the Technical University of Munich, a large European university. This department is currently the largest department at the university with more than 6,000 students."

How to Assign Scarce Resources Without Money:  Designing Information Systems that are Efficient, Truthful, and (Pretty) Fair
Martin Bichler, Alexander Hammerl, Stefan Waldherr,  Thayer Morrill
INFORMATION SYSTEMS RESEARCH, forthcoming.


"Matching with preferences has great potential to coordinate the efficient allocation of scarce resources in organizations when monetary transfers are not available, and thus can provide a powerful design principle for information systems. Unfortunately, it is well-known that it is impossible to combine all three properties of truthfulness, efficiency, and fairness (i.e. envy-freeness) in matching with preferences. Established mechanisms are either efficient or envy-free, and the efficiency loss in envy-free mechanisms is substantial. We focus on a widespread representative of a matching problem: course assignment where students have preferences for courses and organizers have priorities over students. An important feature in course assignment is that a course has both a maximum capacity and a minimum required quota. This is also a requirement in many other matching applications such as school choice, hospital-residents matching, or the assignment of workers to jobs. We introduce RESPCT, a mechanism that respects minimum quotas and is truthful, efficient, and has
low levels of envy. The reduction in envy is significant and is due to two remarkably effective heuristics. We follow a design science approach and provide analytical and experimental results based on field data from a large-scale course assignment application. These results have led to a policy change and the proposed assignment system is now being used to match hundreds of students every semester."

Tuesday, February 18, 2020

Market design and algorithmic criminal justice--by Jung, Kannan, Lee, Pai, Roth and Vohra

When fairness isn't your only goal, your other goals may help you choose among competing definitions of fairness.

Fair Prediction with Endogenous Behavior
Christopher Jung, Sampath Kannan, Changhwa Lee, Mallesh M. Pai, Aaron Roth,and Rakesh Vohra
February 17, 2020

Abstract: There  is  increasing  regulatory  interest  in  whether  machine  learning  algorithms  deployed  in  consequential domains (e.g.  in criminal justice) treat different demographic groups “fairly.”  However, there are several proposed notions of fairness, typically mutually incompatible.  Using criminal justice as an example,  we study a model in which society chooses an incarceration rule.  Agents of different demographic groups differ in their outside options (e.g.  opportunity for legal employment) and decide whether to commit crimes.  We show that equalizing type I and type II errors across groups is consistent with the goal of minimizing the overall crime rate; other popular notions of fairness are not.
*********

And here's a blog post about the paper by one of the authors:

Fair Prediction with Endogenous Behavior
Can Game Theory Help Us Choose Among Fairness Constraints?

"...The crime-minimizing solution is the one that sets different thresholds on posterior probabilities (i.e. uniform thresholds on signals) so as to equalize false positive rates and false negative rates. In other words, to minimize crime, society should explicitly commit to not conditioning on group membership, even when group membership is statistically informative for the goal of predicting crime.

"Why? Its because although using demographic information is statistically informative for the goal of predicting crime when base rates differ, it is not something that is under the control of individuals --- they can control their own choices, but not what group they were born into. And making decisions about individuals using information that is not under their control has the effect of distorting their dis-incentive to commit crime --- it ends up providing less of a dis-incentive to individuals from the higher crime group (since they are more likely to be wrongly incarcerated even if they don't commit a crime). And because in our model people are rational actors, minimizing crime is all about managing incentives. "

Monday, December 23, 2019

"The Ethical Algorithm" by Michael Kearns and Aaron Roth (book talk at Google)

Here's a talk about "The Ethical Algorithm--The Science of Socially Aware Algorithm Design"
by Michael Kearns and Aaron Roth.


IMHO it would make a fine last minute holiday gift for those interested in econ and market design as well as for fans of computer science and algorithms:) 

Saturday, September 28, 2019

Automatic algorithmic affirmative action, by Ashesh Rambachan and Jonathan Roth

There's been some justified concern that algorithms that make predictions and choices based on previous choices made by humans might replicate the human biases embedded in the historic data.  Below is a paper that points out that the opposite effect could happen as well.

As explained here: "Imagine a college that has historically admitted students using (biased) admissions officers, but switches to an algorithm trained on data for their past students. If the admissions officers unfairly set a higher bar for people from group A, then assuming student performance is fairly measured once students arrive on campus, students from group A will appear to be stronger than students from group B. The learned model will therefore tend to favor students from group A, in effect raising the bar for students from group B."*

Here's the paper itself, and its abstract:

Bias In, Bias Out? Evaluating the Folk Wisdom
Ashesh Rambachan, Jonathan Roth

Abstract: We evaluate the folk wisdom that algorithms trained on data produced by biased human decision-makers necessarily reflect this bias. We consider a setting where training labels are only generated if a biased decision-maker takes a particular action, and so bias arises due to selection into the training data. In our baseline model, the more biased the decision-maker is toward a group, the more the algorithm favors that group. We refer to this phenomenon as "algorithmic affirmative action." We then clarify the conditions that give rise to algorithmic affirmative action. Whether a prediction algorithm reverses or inherits bias depends critically on how the decision-maker affects the training data as well as the label used in training. We illustrate our main theoretical results in a simulation study applied to the New York City Stop, Question and Frisk dataset.
**********

* I'm reminded of the saying "To get the same reward as a man, a woman has to be twice as good.  Fortunately that's not hard..."

Tuesday, April 23, 2019

Ethical algoritms: a recent talk and a forthcoming book

Increasingly, algorithms are decision makers. Here's a recent talk, and a book forthcoming in October, about what we might mean by ethical decision making by algorithms.




And here's the forthcoming book:
 The Ethical Algorithm: The Science of Socially Aware Algorithm Design Hardcover – November 1, 2019
by Michael Kearns (Author), Aaron Roth  (Author)

Sunday, March 3, 2019

Impossibility Results in Fairness (from Adventures in Computation)

Here are excerpts from a post at Adventures in Computation that will be of interest to market designers:

Impossibility Results in Fairness as Bayesian Inference

"One of the most striking results about fairness in machine learning is the impossibility result that Alexandra Chouldechova, and separately Jon Kleinberg, Sendhil Mullainathan, and Manish Raghavan discovered a few years ago. These papers say something very crisp. I'll focus here on the binary classification setting that Alex studies because it is much simpler. There are (at least) three reasonable properties you would want your "fair" classifiers to have. They are:
  1. False-Positive Rate Balance: The rate at which your classifier makes errors in the positive direction (i.e. labels negative examples positive) should be the same across groups.
  2. False-Negative Rate Balance:  The rate at which your classifier makes errors in the negative direction (i.e. labels positive examples negative) should be the same across groups.
  3. Predictive Parity: The statistical "meaning" of a positive classification should be the same across groups (we'll be more specific about what this means in a moment)
What Chouldechova and KMR show is that if you want all three, you are out of luck --- unless you are in one of two very unlikely situations: Either you have a perfect classifier that never errs, or the base rate is exactly the same for both populations --- i.e. both populations have exactly the same frequency of positive examples. If you don't find yourself in one of these two unusual situations, then you have to give up on properties 1, 2, or 3.
...
"So why is this result true? The proof in Alex's paper can't be made simpler --- its already a one liner, following from an algebraic identity. But the first time I read it I didn't have a great intuition for why it held. Viewing the statement through the lens of Bayesian inference made the result very intuitive (at least for me). With this viewpoint, all the impossibility result is saying is: "If you have different priors about some event (say that a released inmate will go on to commit a crime) for two different populations, and you receive evidence of the same strength for both populations, then you will have different posteriors as well". This is now bordering on obvious --- because your posterior belief about an event is a combination of your prior belief and the new evidence you have received, weighted by the strength of that evidence.  "
...
"So the mathematical fact is simple --- but its implications remain deep. It means we have to choose between equalizing a decision maker's posterior about the label of an individual, or providing an equally accurate signal about each individual, and that we cannot have both. Unfortunately, living without either one of these conditions can lead to real harm."

Thursday, January 31, 2019

Understanding and misunderstanding algorithmic bias

Adventures in Computation explains a recent political discussion:

Algorithmic Unfairness Without Any Bias Baked In
"Discussion of (un)fairness in machine learning hit mainstream political discourse this week, when Representative Alexandria Ocasio-Cortez discussed the possibility of algorithmic bias, and was clumsily "called out" by Ryan Saavedra on twitter.
...
"Bias in the data is certainly a problem, especially when labels are gathered by human beings. But its far from being the only problem. In this post, I want to walk through a very simple example in which the algorithm designer is being entirely reasonable, there are no human beings injecting bias into the labels, and yet the resulting outcome is "unfair".

Wednesday, October 17, 2018

The FATE of technology at Penn, today

Fairness, Accountability, Transparency and Ethics are topics on which economics and computer science intersect each other , and some other fields as well:

The “FATE” of Technology: Fairness, Accountability, Transparency and Ethics

October 17, 2018 | Glandt Forum, Singh Center for Nanotechnology
Wednesday, October 17th, 2018
Glandt Forum, Singh Center for Nanotechnology
Celebrating Five Years of the Warren Center
Reception to follow

1:00 pm: Welcome and brief remarks
Michael Kearns, Professor and National Center Chair of Computer and Information Science
Rakesh VohraGeorge A. Weiss and Lydia Bravo Weiss University Professor
1:10 pm: Aaron Roth, Class of 1940 Bicentennial Term Associate Professor of Computer and Information Science
“Ethical Algorithms”
1:30 pm: Annie LiangAssistant Professor of Economics
“Predicting and Understanding Human Behaviors through Machine Learning”
1:50 pm: Sandra González-BailónAssociate Professor at the Annenberg School for Communication
“Digital Technologies and Access to News”
2:15 pm: Break
2:30 pm: Panel with members of the Warren Center
Emily Falk, Associate Professor of Communication, Psychology, and Marketing
Junhyong Kim, Patricia M. Williams Term Professor of Biology
Konrad Kording, Penn Integrates Knowledge University Professor of Neuroscience and Bioengineering
3:00 pm: Keynote: Matthew SalganikProfessor of Sociology, Princeton University
“The Fragile Families Challenge”
4:00 pm: Michael Kearns, Professor and National Center Chair of Computer and Information Science
“The AlgoWatch Initiative”
4:30 pm: Concluding remarks by Vijay Kumar, Nemirovsky Family Dean

Tuesday, September 11, 2018

Kidney exchange and computation

Here's an article on how computation--broadly characterized as artificial intelligence--has changed kidney transplantation. It gives some historical background on hard decisions, going back to the first dialysis machines and coming forward to kidney exchange, and has some discussion of  fairness

How AI changed organ donation in the US
By Corinne Purtill

"Today, multiple US hospitals run their own paired kidney donation programs. There are also three larger US exchanges that organize kidney chains across hospitals: the United Network for Organ Sharing, the National Kidney Registry, and the Alliance for Paired Kidney Donation. National exchanges are in place in the UK, Canada, and the Netherlands, and paired donations have taken place in hospitals from India to South Africa.
...
"Given the dearth of public education on what “artificial intelligence” actually means, hospitals and exchanges are wary of patients misconstruing the role algorithms play in identifying potential matches, perhaps fearing conjuring images of robots coldly issuing life-or-death edicts.

Machines currently do not decide which kidneys go where. Humans do that. The algorithms in place today can do the math more reliably and at greater scale than humans can, and implement the judgments humans have already made, but they don’t have a contextual understanding of why they are being asked to perform a calculation in the first place.
...
“In economics we talk about impossibility theorems. There are things you might want that are not possible to get,” Roth says. “When you’re allocating scarce resources, you can’t give a kidney to one person without failing to give it to someone else…. Computers will not lift the burden from humans in every respect.”